m_forward_density_output and m_forward_rgb_output are both 16, why not 1 and 3 as the dimension.

[neilgogogo]

Hi, I have a question that why the dimension of the m_forward_density_output and m_forward_rgb_output are both 16. As for the density, only the first element in the 16 elements is used and as for the rgb_output, only the first 3 elements are used every 16 elements. So, why not switch the dimension into 1 and 3 respectively to save the memory?

Many thanks.

[Tom94]

The hardware matrix multipliers of modern GPUs operate on blocks of data with fixed size. The underlying FullyFusedMLP, for example, uses 16x16x16 tensor core functionality. Ergo, outputs must be padded to 16 dimensions in order to use the acceleration.